In a study of 1,053 professional developers reviewing hidden dangerous commands in real coding sessions, human reviewers caught 13.6% while auto mode caught 89%. Starting tomorrow, auto mode becomes the default permission mode in Claude Code for Pro, Max, and Team plans.
Follow-up to the announcement that auto mode is becoming the default permission mode for Pro, Max, and Team plans, linking to further reading on the research behind the change.
A user raised concerns that Anthropic's text watermarking might subtly alter verbatim quotes, such as legal documents, to embed the watermark. The answer is an unambiguous no; quoted source material will not be changed. Technical details are still being prepared for sharing.
Named Claude Code sessions can now DM each other directly from the CLI. Launch sessions with --name flags (e.g. "backend" and "frontend") and send messages between them inline, which apparently works quite well for coordinating multi-agent workflows.
A user's backend and frontend Claude Code sessions spent two hours arguing about snake_case vs camelCase before the process was killed. The team member's only response: were they making valid arguments?
Responding to concerns that steganographic text watermarking might degrade model outputs by steering away from highest-probability tokens, the team says there is no quality impact. More technical details on how the watermarking works are coming.
